Fix incorrect NEON native fp16 multiplication.
diff --git a/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h b/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
index 00bdb9b..b97a090 100644
--- a/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/arch/NEON/GeneralBlockPanelKernel.h
@@ -2,7 +2,7 @@
namespace Eigen {
namespace internal {
-
+
#if EIGEN_ARCH_ARM && EIGEN_COMP_CLANG
// Clang seems to excessively spill registers in the GEBP kernel on 32-bit arm.
@@ -218,7 +218,9 @@
EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
{
- loadRhs(b,dest);
+ // If LHS is a Packet8h, we cannot correctly mimic a ploadquad of the RHS
+ // using a single scalar value.
+ eigen_assert(false && "Cannot loadRhsQuad for a scalar RHS.");
}
EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& /*tmp*/, const FixedInt<0>&) const
diff --git a/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
index bfae650..4a6cef5 100644
--- a/Eigen/src/Core/products/GeneralBlockPanelKernel.h
+++ b/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -751,6 +751,9 @@
template<typename Packet> struct unpacket_traits<DoublePacket<Packet> > {
typedef DoublePacket<typename unpacket_traits<Packet>::half> half;
+ enum{
+ size = 2 * unpacket_traits<Packet>::size
+ };
};
// template<typename Packet>
// DoublePacket<Packet> pmadd(const DoublePacket<Packet> &a, const DoublePacket<Packet> &b)
@@ -2490,7 +2493,13 @@
// nr (which is currently 4) for the return type.
const int SResPacketHalfSize = unpacket_traits<typename unpacket_traits<SResPacket>::half>::size;
const int SResPacketQuarterSize = unpacket_traits<typename unpacket_traits<typename unpacket_traits<SResPacket>::half>::half>::size;
- if ((SwappedTraits::LhsProgress % 4) == 0 &&
+ // The following code assumes we can load SRhsPacket in such a way that
+ // it multiplies blocks of 4 elements in SLhsPacket. This is not the
+ // case for some customized kernels (i.e. NEON fp16). If the assumption
+ // fails, drop down to the scalar path.
+ constexpr bool kCanLoadSRhsQuad = (unpacket_traits<SLhsPacket>::size < 4) || (unpacket_traits<SRhsPacket>::size % (unpacket_traits<SLhsPacket>::size / 4)) == 0;
+ if (kCanLoadSRhsQuad &&
+ (SwappedTraits::LhsProgress % 4) == 0 &&
(SwappedTraits::LhsProgress<=16) &&
(SwappedTraits::LhsProgress!=8 || SResPacketHalfSize==nr) &&
(SwappedTraits::LhsProgress!=16 || SResPacketQuarterSize==nr))
diff --git a/test/product_small.cpp b/test/product_small.cpp
index 02d5f35..a3a04dc 100644
--- a/test/product_small.cpp
+++ b/test/product_small.cpp
@@ -281,6 +281,25 @@
}
}
+template<typename T>
+void product_sweep(int max_m, int max_k, int max_n) {
+ using Matrix = Eigen::Matrix<T, Eigen::Dynamic, Eigen::Dynamic>;
+ for (int m = 1; m < max_m; ++m) {
+ for (int n = 1; n < max_n; ++n) {
+ Matrix C = Matrix::Zero(m, n);
+ Matrix Cref = Matrix::Zero(m, n);
+ for (int k = 1; k < max_k; ++k) {
+ Matrix A = Matrix::Random(m, k);
+ Matrix B = Matrix::Random(k, n);
+ C = A * B;
+ Cref.setZero();
+ ref_prod(Cref, A, B);
+ VERIFY_IS_APPROX(C, Cref);
+ }
+ }
+ }
+}
+
EIGEN_DECLARE_TEST(product_small)
{
for(int i = 0; i < g_repeat; i++) {
@@ -290,7 +309,7 @@
CALL_SUBTEST_3( product(Matrix3d()) );
CALL_SUBTEST_4( product(Matrix4d()) );
CALL_SUBTEST_5( product(Matrix4f()) );
- CALL_SUBTEST_50( product(Matrix<bfloat16, 3, 2>()) );
+ CALL_SUBTEST_10( product(Matrix<bfloat16, 3, 2>()) );
CALL_SUBTEST_6( product1x1<0>() );
CALL_SUBTEST_11( test_lazy_l1<float>() );
@@ -317,6 +336,12 @@
CALL_SUBTEST_6( bug_1311<5>() );
CALL_SUBTEST_9( test_dynamic_bool() );
+
+ // Commonly specialized vectorized types.
+ CALL_SUBTEST_50( product_sweep<float>(10, 10, 10) );
+ CALL_SUBTEST_51( product_sweep<double>(10, 10, 10) );
+ CALL_SUBTEST_52( product_sweep<Eigen::half>(10, 10, 10) );
+ CALL_SUBTEST_53( product_sweep<Eigen::bfloat16>(10, 10, 10) );
}
CALL_SUBTEST_6( product_small_regressions<0>() );
